19th IEEE International Conference on Tools with Artificial Intelligence
Extraction-Based Single-Document Summarization Using
Random Indexing
Niladri Chatterjee
Department of Mathematics
Indian Institute of Technology Delhi
New Delhi, India – 110016
niladri@maths.iitd.ac.in
Abstract
This paper presents a summarization technique for
text documents exploiting the semantic similarity between
sentences to remove the redundancy from the text.
Semantic similarity scores are computed by mapping the
sentences on a semantic space using Random Indexing.
Random Indexing, in comparison with other semantic
space algorithms, presents a computationally efficient
way of implicit dimensionality reduction. It involves
inexpensive vector computations such as addition. It thus
provides an efficient way to compute similarities between
words, sentences and documents. Random Indexing has
been used to compute the semantic similarity scores of
sentences and graph-based ranking algorithms have been
employed to produce an extract of the given text.
1. Introduction
Automatic Text Summarization is an important and
challenging area of Natural Language Processing. The
task of a text summarizer is to produce a synopsis of any
document or a set of documents submitted to it.
Summaries differ in several ways. A summary can be
an extract i.e. certain portions (sentences or phrases) of
the text is lifted and reproduced verbatim, whereas
producing an abstract involves breaking down of the text
into a number of different key ideas, fusion of specific
ideas to get more general ones, and then generation of
new sentences dealing with these new general ideas [1]. A
summary can be of a single document or multiple
documents, generic (author’s perspective) or query
oriented (user specific) [2], indicative (using keywords
1082-3409/07 $25.00 © 2007 IEEE
DOI 10.1109/ICTAI.2007.28
Shiwali Mohan
Division of Instrumentation and Control
Netaji Subhas Institute of Technology
New Delhi, India - 110075
shiwali.mohan@gmail.com
indicating the central topics) or informative (content
laden) [3].
In this work we have focused on producing a generic,
extractive, informative, single document summary
exploiting the semantic similarity of sentences.
2. Previous Work in Extractive Text
Summarization
Various methods have been proposed to achieve
extractive summarization. Most of them are based on
scoring of the sentences. Maximal Marginal Relevance [4]
scores the sentences according to their relevance to the
query, Mutual Reinforcement Principle for Summary
generation [5] uses clustering of sentences to score them
according to how close they are to the central theme. QR
decomposition method [6] scores the sentences using
column pivoting. The sentences can also be scored by
certain predefined features. These features may include
linguistic features and statistical features, such as location,
rhetorical structure [7], presence or absence of certain
syntactic features [8] and presence of proper names, and
statistical measures of term prominence [9].
Rough set based extractive summarization [10] has
been proposed that aims at selecting important sentences
from a given text using rough sets, which has been
traditionally used to discover patterns hidden in data.
Methods using similarity between sentences and
measures of prominence of certain semantic concepts and
relationships [11] to generate an extractive summary have
also been proposed.
Some commercially available extractive summarizers
like Copernic [12] and Word [13] summarizers use
certain statistical algorithms to create a list of important
concepts and hence generate a summary.
448
 We propose to achieve extractive summarization as a
three-step process:
1. Mapping the words and sentences onto a semantic
space (Word Space)
2. Computing similarities between them.
3. Employing the use of graph-based ranking algorithms
[14] to remove the redundant sentences in the text.
The task involves simple mathematical computations,
such as addition of vectors, and thus is far more effective
than other algorithms based on semantic similarities, such
as LSA based summarization that involves expensive
matrix computations.
An environment in linguistics is called a context. A
context of a word can easily be understood as the
linguistic surrounding of the word. As an illustration,
consider the following sentence.
‘A friend in need is a friend indeed’
If we define the context of a focus word as one preceding
and one succeeding word, then the context of ‘need’ is
‘in’ and ‘is’, whereas the context of ‘a’ is ‘is’ and ‘friend’.
To tabulate this context information a co-occurrence
matrix of the following form can be created, in which the
(i,j)th element denotes the number of times word i occurs
in the context of word j within the text.

3. The Word Space Model Co-occurrents

Word a friend in need is indeed
The Word-Space Model [15] is a spatial a 0 2 0 0 1 0
representation of word meaning. The complete vocabulary friend 2 0 1 0 0 1
of any text (containing n words) can be represented in an in 0 1 0 1 0 0
n-dimensional space in which each word occupies a need 0 0 1 0 1 0
specific point in the space and has a vector associated is 1 0 0 1 0 0
with it defining its meaning. indeed 0 1 0 0 0 0
The Word Space Model is based entirely on language
data available. It does not rely on previously compiled
lexicons or databases to represent the meaning. It only Here the context vector for ‘need’ is [ 0 0 1 0 1 0] and
represents what is really there in the current universe of for ‘a’ is [0 2 0 0 1 0]. They effectively sum up the
discourse. When meanings change, disappear or appear in environments (contexts) of the words in question, and can
the data at hand, the model changes accordingly. If an be represented in a six-dimensional geometric space
entirely different set of language data is used, a complete (since the text contains 6 words). A context vector thus
different model of meaning is obtained. obtained can be used to represent the distributional
information of the word into geometrical space. (Note that
3.1. The Word Space Hypotheses this is similar to assigning a unary index vector to ‘is’ ([ 0
0 0 0 1 0 ]) and to ‘in’ ([0 0 1 0 0 0]) and adding them up
The Word Space is so constructed that the following to get the context vector of ‘need’.)
two hypotheses hold true.
3.3. Similarity in Mathematical Terms
1. The Proximity Hypothesis: The words which lie closer
to each other in the word space have similar meanings Context vectors as such do not convey any beneficial
while the words distant in the word space have dissimilar information. They just give the location of the word in the
meanings. word space. To get a clue on ‘how similar the words are
in their meaning’ a similarity measure of the context
2. The Distributional Hypothesis: Once the language data vectors is required.
is obtained, the word space model uses the statistics about Various schemes, such as scalar product of vectors,
the distributional properties of the words. The words Euclidean distance, Minkowski metrics [15], can be used
which are used within similar group of words (i.e. similar to compute similarity between vectors.
context) should be placed nearer to each other. We have used cosine of the angles between the two
vectors x and y to compute normalized vector similarity.
3.2. Context Vectors and Co-occurrence matrices The cosine angle between two vectors x and y is defined
as:
Once the distributional property of a word is obtained,
the next step is to transform the distributional information
into a geometric representation. In other words “The
distribution of an element will be understood as the sum
of all its environments” [16].

449
 The cosine measure corresponds to taking the scalar
product of the vectors and then dividing by their norms.
The cosine measure is the most frequently utilized
similarity metric in word-space research. The advantage
of using cosine metric over other metrics to calculate
similarity is that it provides a fixed measure of similarity,
which ranges from 1 (for identical vectors), to 0 (for
orthogonal vectors) and -1 (for vectors pointing in the
opposite directions). Moreover, it is also comparatively
efficient to compute.
3.4. Problems Associated with Implementing
Word Spaces
The dimension n used to define the word space
corresponding to a text document is equal to the number
of unique words in the document. The number of
dimensions increases as the size of text increases. Thus a
text document containing a few thousands of words will
have a word space of few thousands of dimensions. Thus
computational overhead increases rapidly with the size of
the text. The other problem is of data sparseness. The
majority of cells in co-occurrence matrix constructed
corresponding to the document will be zero. The reason is
that the most of the words in any language appear in
limited context, i.e. the words they co-occur with are very
limited.
The solution to this predicament is to reduce the high
dimensionality of the vectors. A few algorithms attempt to
solve this problem by dimensionality reduction. One of
the simplest ways is to remove words belonging to certain
grammatical classes. Other way could be employing
Latent Semantic Analysis [17]. We have used Random
Indexing [18] to address the problem of high
dimensionality.
4. Random Indexing
Random Indexing (RI) [18] is based on Pentti
Kanerva's [19] work on sparse distributed memory.
Random Indexing was developed to tackle the problem of
high dimensionality in word space model. While
dimensionality reduction does make the resulting lower-
dimensional context vectors easier to compute with, it
does not solve the problem of initially having to collect a
potentially huge co-occurrence matrix. Even
implementations that use powerful dimensionality
reduction, such as SVD [17], need to initially collect the
words-by-documents or words-by-words co-occurrence
matrix. RI removes the need for the huge co-occurrence
matrix. Instead of first collecting co-occurrences in a co-
occurrence matrix and then extracting context vectors
from it, RI incrementally accumulates context vectors,
450
which can then, if needed, be assembled into a co-
occurrence matrix.
4.1. RI Algorithm
Random Indexing accumulates context vectors in a
two step process:
1. Each word in the text is assigned a unique and
randomly generated vector called the index vector. The
index vectors are sparse and high dimensional and ternary
(i.e. 1, -1, 0). Each word is also assigned an initially
empty context vector which has the same dimensionality
(r) as the index vector.
2. The context vectors are then accumulated by advancing
through the text one word taken at a time, and adding the
context's index vector to the focus word's context vector.
When the entire data has been processed, the r-
dimensional context vectors are effectively the sum of the
words' contexts.
For illustration we can again take the example of the
sentence
‘A friend in need is a friend indeed’
Let the dimension r of the index vector be 10 for
illustration purposes. The context is defined as one
preceding and one succeeding word.
Let ‘friend’ be assigned a random index vector:
[0 0 0 1 0 0 0 0 -1 0 ]
and ‘need’ be assigned a random index vector:
[0 1 0 0 -1 0 0 0 0 0]
Then to compute the context vector of ‘in’ we need to
sum up the index vector of its context. Since the context is
defined as one preceding and one succeeding word, the
context of ‘in’ is ‘friend’ and ‘need’. We sum up their
index vectors to get the context vector of ‘in’.
[0 1 0 1 -1 0 0 0 -1 0]
If a co-occurrence matrix has to be constructed, r-
dimensional context vectors can be collected into a matrix
of order w x r, where w is the number of unique word
types, and r is the chosen dimensionality of for each word.
Note that this is similar to constructing an n-
dimensional unary context vector which has a single 1 in
different positions for different words and n is the number
of distinct words. Mathematically, these n dimensional
unary vectors are orthogonal, whereas the r-dimensional
random index vectors are nearly orthogonal. There are
many more nearly orthogonal than truly orthogonal
directions in a high-dimensional space [18]. Choosing
random indexing is an advantageous tradeoff between the
number of dimensions and orthogonality, as the r-
dimensional random index vectors can be seen as
approximations of the n-dimensional unary vectors.
 Observe that both the unary vectors and the random
index vectors assigned to the words construct the word
space. The context vectors computed on the language data
are used in mapping the words onto the word space. In
our work we used Random Indexing because of the
advantages discussed below.
4.2. Advantages of Random Indexing
Compared to other word space methodologies Random
Indexing approach is unique in the following three ways:
First, it is an incremental method, which means that
the context vectors can be used for similarity
computations even after just a few examples have been
encountered. By contrast, most other word space methods
require the entire data to be sampled before similarity
computations can be performed.
Second, it uses fixed dimensionality, which means that
new data do not increase the dimensionality of the
vectors. Increasing dimensionality can lead to significant
scalability problems in other word space methods.
Third, it uses implicit dimension reduction, since the
fixed dimensionality is much lower than the number of
words in the data. This leads to a significant gain in
processing time and memory consumption as compared to
word space methods that employ computationally
expensive dimension reduction algorithms.
4.3. Assigning Semantic Vectors to Documents
The average term vector can be considered as
the central theme of the document and is computed as:
where n is the number of distinct words in the document.
While we compute the semantic vectors for the
sentences we subtract from the context vectors
of the words of the sentence to remove the bias from the
system [21]. The semantic vector of a sentence is thus
computed as:
where, n is the number of words in the focus sentence and
th
i refers to the i word of the sentence and is the
corresponding context vector.
Note that subtracting the mean vector reduces the
magnitude of those term vectors which are close in
direction to the mean vector, and increases the magnitude
of term vectors which are most nearly opposite in
direction from the mean vector. Thus the words which
occur very commonly in a text, such as the auxiliary verbs
and articles, will have little influence on the sentence
vector so produced. Typically, these words do not have
451
any definitive pattern about the words they co-occur with.
Further, the terms whose distribution is most distinctive
will be given the most weight.
5. The Experimental Setup
Our experimental data set consists of fifteen
documents containing 200 to 300 words each. The
processing of each document to generate a summary has
been carried out as follows:
5.1. Mapping of Words onto the Word Space
Each word in the document was initially assigned a
unique randomly generated index vector of the dimension
100 with ternary values (1, -1, 0). This provided an
implicit dimensionality reduction of around 50%. The
index vectors were so constructed that each vector of 100
units contained two randomly placed 1 and two randomly
placed -1s, rest of the units were assigned 0 value. Each
word was also assigned an initially empty context vector
of dimension 100. The dimensions r assigned to the words
depend upon the number of unique words in the text.
Since our test data consisted of small paragraphs of 200-
300 words each the vector of dimensions 100 sufficed. If
larger texts containing thousands of word are to be
summarized larger dimensional vectors have to be
employed.
We defined the context of a word as two words on
either side. Thus a 2x2 sliding window was used to
accumulate the context vector of the focus word. The
context of a given word was also restricted in one
sentence, i.e. across sentence windows were not
considered. In case where the window extended in the
preceding or the succeeding sentence, a unidirectional
window was used. There is fair evidence supporting the
use of small context window. Kaplan [22] conducted
various experiments with people in which they
successfully guessed the meaning of a word if two words
on either side of it were also provided. Experiments
conducted at SICS, Sweden [23] also indicate that a
narrow context window is preferable for acquiring
semantic information. The above observation prompted us
to use a 2x2 window. The window can be weighted as
well to give greater importance to the words lying closer
to the focus word. For example, the weight vector [0.5 1 0
1 0.5] suggests that the words adjacent to the focus word
are given the weight 1 and the words at distance 2 are
assigned a weight of 0.5. In our experiments we have used
the above mentioned weights for computing the context
vectors.
5.2. Mapping of Sentences onto the Word Space
Once all the context vectors have been accumulated,
semantic vectors for the sentences were computed. A
mean vector was calculated from the context vectors of all
the words in the text. This vector was subtracted from the
context vectors of the word appearing in the sentence, the
resultants were summed up and averaged to compute the
semantic vector of the sentence.
5.3. Construction of Completely Connected
Undirected Graph
We constructed a weighted, completely connected,
undirected graph from the text, wherein each sentence is
represented by a node in the graph. The edge joining node
i and node j is associated with a weight wij signifying the
similarity between the sentence i and sentence j.
5.3.1. Assigning the Node Weights
The weight of each node was determined by
calculating the ‘relevance’ of each sentence in the text.
For this purpose, we identified the index words [9] of the
document. An index word is a word which has a
frequency higher than a predetermined lower cutoff and
does not belong to the grammatical classes like articles,
prepositions and auxiliary verbs. All the index words were
assigned a weight which was calculated by dividing the
number of occurrence of the word by total number of
distinct words in the text.
For example in the text containing 6 distinct words:
“A friend in need is a friend indeed” the word ‘friend’
occurs twice. Thus it is assigned a weight of 2/6 = 0.333.
A 1-dimensional vector was allocated to each
sentence, with length equal to the number of index words
and each element referring to an index word. The value of
that element was determined by multiplying the number of
times the index word occurred in the sentence by its
weight.
An average document vector was calculated by
averaging the context vectors of all its sentences. Cosine
similarity of each of the sentence with the document
vector was calculated. The value thus obtained was
assigned to the respective node and will be called the
node weight.
5.3.2. Assigning the Edge Weights
The edges between the nodes are weighted by the
similarity scores between the participating nodes
(sentences). The similarity was computed by determining
the cosine metric between the sentence vectors.
452
5.4 Calculating Weights of the Sentences and
Generating Summary
Once the graph is constructed, our aim is to get rid of
the redundant information in the text by removing the
sentences of less importance. To achieve this, the
sentences are ranked by applying some graph-based
ranking algorithms. Various graph-based ranking
algorithms are available in literature. The one that we
have used for this work is weighted PageRank algorithm
[24].
Weighted PageRank Algorithm
Let G = (V, E) be a directed graph with the set of
vertices V and set of edges E, where E is a subset of VxV.
For a given vertex Vi, let In(Vi) be the set of vertices that
point to it (predecessors), and let Out(Vi) be the set of
vertices that vertex Vi points to (successors). Then the
new node weight assigned by PageRank ranking algorithm
after one iteration is:
This computation is carried out on all the nodes in
succession iteratively until node weights converge. We set
the value 0.85 to the factor d as per the recommendation
in [25]. We have applied this algorithm on undirected
graphs constructed for each text by considering In(Vi) =
Out(Vi) for all nodes.
The node weights converge in such a way that weights
of the important sentences are highly increased, while
those of the redundant sentences do not increase in same
proportion. Once the weights stabilize, the sentences are
sorted in descending order of their weights. The top few
sentences are selected to generate the summaries.
6. Results
We have run the experiments on different texts and
computed extracts at 10%, 25%, 50% levels. We
compared the results with manual summaries created by
experts and also the summaries generated by some
commercially available summarizers namely Copernic
and Word. Below we show a sample text and its
summaries at different percentage levels generated by our
scheme. We also show the summaries generated by
Copernic and Word at different levels and also the expert
generated summary.
Consider the text given below as our sample
document, which is a ten-sentence long document. As
mentioned earlier, for evaluation purpose we have used
documents which are 200 to 300 words long.
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
This configuration can only occur during a new moon, when the Sun
and Moon are in conjunction as seen from the Earth.
In ancient times, and in some cultures today, solar eclipses are
attributed to mythical properties.
Total solar eclipses can be frightening events for people unaware of
their astronomical nature, as the Sun suddenly disappears in the middle
of the day and the sky darkens in a matter of minutes.
However, the spiritual attribution of solar eclipses is now largely
disregarded.
Total solar eclipses are very rare events for any given place on Earth
because totality is only seen where the Moon's umbra touches the
Earth's surface.
A total solar eclipse is a spectacular natural phenomenon and many
people consider travel to remote locations in order to observe one.
The 1999 total eclipse in Europe, said by some to be the most-watched
eclipse in human history, helped to increase public awareness of the
phenomenon.
This was illustrated by the number of people willing to make the trip to
witness the 2005 annular eclipse and the 2006 total eclipse.
The next solar eclipse takes place on September 11, 2007, while the
next total solar eclipse will occur on August 1, 2008.
The sentences selected by experts manually to create a
summary are:
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
This configuration can only occur during a new moon, when the Sun
and Moon are in conjunction as seen from the Earth.
Total solar eclipses are very rare events for any given place on Earth
because totality is only seen where the Moon's umbra touches the
Earth's surface.
A total solar eclipse is a spectacular natural phenomenon and many
people consider travel to remote locations in order to observe one.
The 1999 total eclipse in Europe, said by some to be the most-watched
eclipse in human history, helped to increase public awareness of the
phenomenon.
The summary generated by our summarizer:
10% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
25% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
This configuration can only occur during a new moon, when the Sun
and Moon are in conjunction as seen from the Earth.
Total solar eclipses can be frightening events for people unaware of
their astronomical nature, as the Sun suddenly disappears in the middle
of the day and the sky darkens in a matter of minutes.
50% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
This configuration can only occur during a new moon, when the Sun
and Moon are in conjunction as seen from the Earth.
Total solar eclipses can be frightening events for people unaware of
their astronomical nature, as the Sun suddenly disappears in the middle
of the day and the sky darkens in a matter of minutes.
Total solar eclipses are very rare events for any given place on Earth
because totality is only seen where the Moon's umbra touches the
Earth's surface.
The 1999 total eclipse in Europe, said by some to be the most-watched
eclipse in human history, helped to increase public awareness of the
phenomenon.
453
The summary generated by Copernic:
10% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
25% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
A total solar eclipse is a spectacular natural phenomenon and many
people consider travel to remote locations in order to observe one.
50% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
total solar eclipses can be frightening events for people unaware of their
astronomical nature, as the Sun suddenly disappears in the middle of the
day and the sky darkens in a matter of minutes.
A total solar eclipse is a spectacular natural phenomenon and many
people consider travel to remote locations in order to observe one.
The 1999 total eclipse in Europe, said by some to be the most-watched
eclipse in human history, helped to increase public awareness of the
phenomenon.
The summary generated by Word:
10% summary
The next solar eclipse takes place on September 11, 2007, while the
next total solar eclipse will occur on August 1, 2008.
25% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
A total solar eclipse is a spectacular natural phenomenon and many
people consider travel to remote locations in order to observe one.
The next solar eclipse takes place on September 11, 2007, while the
next total solar eclipse will occur on August 1, 2008.
50% summary
A solar eclipse occurs when the Moon passes between Earth and the
Sun, thereby totally or partially obscuring Earth's view of the Sun.
A total solar eclipse is a spectacular natural phenomenon and many
people consider travel to remote locations in order to observe one.
The 1999 total eclipse in Europe, said by some to be the most-watched
eclipse in human history, helped to increase public awareness of the
phenomenon.
This was illustrated by the number of people willing to make the trip to
witness the 2005 annular eclipse and the 2006 total eclipse.
The next solar eclipse takes place on September 11, 2007, while the
next total solar eclipse will occur on August 1, 2008.
For larger texts, we used Precision, Recall and F, widely
used in Information Retrieval [26] for evaluating our
results. For each document an extract done manually by
experts has been considered as the reference summary
(denoted by Sref). We then compare the candidate
summary (denoted by Scand) with the reference summary
and compute the precision, recall and F values as follows:
We also compute the precision, recall and F values for the
summaries generated by Copernic [12] and Word
summarizer [14] by comparing them with Sref. Finally we
compare the p, r, F values corresponding to our
summarizer with these values. The values obtained for
fifteen documents have been tabulated in Table 1.
 Text Our Approach Copernic Word
Number
p r F p r F p r F
1 10% 1.000 0.166 0.284 1.000 0.083 0.154 1.000 .250 0.400
25% 0.800 0.333 0.424 0.500 0.166 0.252 0.800 .333 0.424
2 10% 1.000 0.444 0.444 1.000 0.140 0.250 0.500 0.142 0.222
25% 1.000 0.429 0.601 1.000 0.429 0.545 0.333 0.142 0.200
3 10% 0.500 .125 0.200 0.500 0.125 0.200 0.500 0.125 0.200
25% 0.600 .375 0.462 0.750 0.375 0.500 0.600 0.375 0.462
4 10% 1.000 0.200 0.333 1.000 0.200 0.333 0 0 N.A.
25% 1.000 0.400 0.570 0.666 0.400 0.498 0.666 0.400 0.498
5 10% 1.000 0.143 0.249 1.000 0.143 0.249 0.500 0.143 0.222
25% 0.750 0.426 0.545 0.666 0.286 0.400 0.500 0.286 0.329
6 10% 1.00 0.300 0.451 1.000 0.200 0.333 0.500 0.100 0.167
25% 0.833 0.500 0.625 0.666 0.300 0.400 0.400 0.200 0.267
7 10% 1.000 0.222 0.364 1.000 0.222 0.364 0.500 0.111 0.181
25% 0.200 0.444 0.552 0.750 0.333 0.458 0.400 0.400 0.282
8 10% 1.000 0.250 0.400 1.000 0.250 0.400 0 0 N.A.
25% 1.000 0.500 0.666 1.000 0.500 0.666 0.500 0.250 0.400
9 10% 0.750 0.200 0.315 0.666 0.133 0.221 0.500 0.133 0.210
25% 0.875 0.466 0.608 0.857 0.400 0.545 0.625 0.333 0.434
10 10% 0.666 0.200 0.307 1.000 0.200 0.333 0.500 0.200 0.285
25% 0.833 0.500 0.625 0.800 0.400 0.533 0.714 0.500 0.588
11 10% 1.000 0.166 0.285 1.000 0.166 0.284 0.666 0.166 0.265
25% 0.857 0.500 0.632 0.666 0.333 0.444 0.875 0.583 0.699
12 10% 0.666 0.125 0.211 1.000 0.125 0.222 0.666 0.125 0.210
25% 0.875 0.438 0.583 0.750 0.375 0.500 0.750 0.375 0.500
13 10% 1.000 0.222 0.363 1.000 0.222 0.363 0.666 0.222 0.333
25% 0.800 0.444 0.571 0.750 0.333 0.461 0.600 0.333 0.428
14 10% 1.000 0.182 0.305 1.000 0.182 0.308 0.600 0.182 0.279
25% 0.833 0.454 0.593 0.800 0.364 0.499 0.600 0.364 0.453
15 10% 1.000 0.230 0.373 1.000 0.230 0.373 0.666 0.154 0.249
25% 0.857 0.461 0.599 0.666 0.307 0.419 0.714 0.384 0.499

Table 1: Results of Our summariser, Copernic and Word Summarizer

The observations clearly indicate that the summaries
generated by our method are closer to the human
generated summaries that the summaries produced by
Copernic and Word summarizers at 10% and 25% level
in almost all the text cases. At 50% level too we
obtained better results compared to Copernic and Word.
However, limitation of space precludes us to show the
figures in this table.
6. Conclusions and Future Scope
In this paper we have proposed a summarization
technique which involves mapping of the words and
sentences onto a semantic space and exploiting their
similarities to remove the less important sentences
containing redundant information. The problem of high
dimensionality of the semantic space corresponding to
the text has been tackled by employing Random
Indexing which is less expensive in computations and
memory consumption compared to other dimensionality
reduction approaches. The approach gives better results
than commercially available summarizers namely
Copernic and Word Summarizer.
In future we plan to include a training algorithm
using Random Indexing which will construct the Word
Space on a previously compiled text database and then
to use it for summarization purposes so as to resolve the
ambiguities, such as polysemy, more efficiently.
We observed some abruptness in the summaries
generated by our method. We plan to smoothen out this
abruptness by constructing Stiener trees of the graphs
constructed corresponding to the text.
In our present evaluation we have used measures like
precision, recall and F which are used primarily in the
context of information retrieval. In future we intend to
use more summarization-specific techniques, e.g.
ROUGE [27] to measure the efficacy of our scheme.

454
 Text summarization is an important challenge in
present day context for huge volumes of text are being
produced every day. We expect that the proposed
approach will paves the way for developing an efficient
AI tool for text summarization.
7. References
[1] Inderjeet Mani, “Advances in Automatic Text
Summarization”, MIT Press, Cambridge, MA, USA, 1999.
[2] Jade Goldstein, Mark Kantrowitz, Vibhu Mittal, and Jaime
Carbonell, “Summarizing text documents:sentence selection
and evaluation metrics”, ACM SIGIR, 1999, pp 121–128.
[3] E.H. Hovy and C.Y. Lin, “Automated Text Summarization
in SUMMARIST”, Proceedings of the Workshop on
Intelligent Text Summarization, ACL/EACL-97. Madrid,
Spain, 1997.
[4] J. Carbonell and J. Goldstein, “The use of MMR, diversity-
based reranking for reordering documents and producing
summaries,” ACM SIGIR, 1998, pp. 335–336.
[5] Zha Hongyuan, “Generic Summarization and Key phrase
Extraction Using Mutual Reinforcement Principle and
Sentence Clustering”, ACM, 2002.
. IJCNN'98 IEEE Service Center, 1999.
[6] John Conroy, Leary Dianne, “Text Summarization via
Hidden Markov Models and Pivoted QR Matrix
Decomposition”, ACM SGIR ,2001.
[7] Daniel Marcu “From discourse structures to text
summaries” In ACL’97/EACL’97 Workshop on Intelligent
Scalable Text Summarization, 1997, pp 82–88.
. [22] A. Kaplan “An experimental study of ambiguity and
[8] J. Pollock and A. Zamora “Automatic abstracting research
at chemical abstracts service”, JCICS, 1975.
[9] Hans P. Luhn, “The automatic creation of literature
abstracts”, IBM J. of R. and D, 1958.
[10] Nidhika Yadav, M.Tech Thesis , Indian Institute of
Technology Delhi, 2007
[11] Yihong Gong and Xin Liu, “Generic text summarization
using relevance measure and latent semantic analysis”,
SIGIR,ACM, 2001, pp 19–25.
[12] Copernic Summarizer Homepage: http://www.copernic
.com /en/products/summarizer.
[13] Rada Mihalcea and Paul Tarau, “An Algorithm for
Language Independent Single and Multiple Document
Summarization”, Proceedings of the International Joint
Conference on Natural Language Processing (IJCNLP),
Korea, October 2005.
455
[14] Word Sumamriser www.microsoft.com/education/
autosummarize.mspx
[15] M. Sahlgren, “The Word-Space Model: Using
distributional analysis to represent syntagmatic and
paradigmatic relations between words in high-dimensional
vector spaces”, Ph.D. dissertation, Department of Linguistics,
Stockholm University, 2006
[16] Z. Harris, Mathematical structures of language,
Interscience Publishers, 1968.
[17] Thomas K Landauer, Peter W. Foltz, Darrell Laham, “An
Introduction to Latent Semantic Analysis”, 45th Annual
Computer Personnel Research Conference – ACM, 2004.
[18] M. Sahlgren, “An Introduction to Random Indexing.
Proceedings of the Methods and Applications of Semantic
Indexing”, Workshop at the 7th International Conference on
Terminology and Knowledge Engineering, TKE, Copenhagen,
Denmark, 2005.
[19] P. Kanerva, Sparse distributed memory, Cambridge, MA,
USA: MIT Press, 1988
[20] S. Kaski, “Dimensionality reduction by random mapping:
Fast similarity computation for clustering”, Proceedings of the
International Joint Conference on Neural Networks,
[21] Derrick Higgins, Jill Burstein, “Sentence similarity
measures for essay coherence”, Proceedings of the 2004
Human language Technology Conference of the North
American chapter of the Association for Computational
Linguistics, Boston, Massachusetts, May 2004
context”, Mechanical Translation, 2(2), 1955.
[23] J. Karlgren, & M. Sahlgren (2001), “From words to
understanding”, Foundations of real-world intelligence ,CSLI
Publications, 2001
[24] Rada Mihalcea, “Graph-based Ranking Algorithms for
Sentence Extraction, Applied to Text Summarization”,
Proceedings of the 42nd Annual Meeting of the Association
for Computational Linguistics, companion volume (ACL
2004), Barcelona, Spain, July 2004.
[25] S. Brin and L. Page, “The anatomy of a large-scale
hypertextual Web search engine”, Computer Networks and
ISDN Systems, 1998.
[26] R.B. Yates, B.R. Neto, Modern Information Retrieval,
Pearson Education, 1999
[27] ROUGE: Recall Oriented Understudy for Gisting
evaluation, http://www.isi.edu/~cyl/ROUGE/